| Message ID | 20240228232253.2875900-3-dw@davidwei.uk (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | 9eb95228a74163e26d338913a5d61e571ea45307 |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | netdevsim: link and forward skbs between ports | expand |
On 29/02/2024 00:22, David Wei wrote: > Forward skbs sent from one netdevsim port to its connected netdevsim > port using dev_forward_skb, in a spirit similar to veth. > > Add a tx_dropped variable to struct netdevsim, tracking the number of > skbs that could not be forwarded using dev_forward_skb(). > > The xmit() function accessing the peer ptr is protected by an RCU read > critical section. The rcu_read_lock() is functionally redundant as since > v5.0 all softirqs are implicitly RCU read critical sections; but it is > useful for human readers. > > If another CPU is concurrently in nsim_destroy(), then it will first set > the peer ptr to NULL. This does not affect any existing readers that > dereferenced a non-NULL peer. Then, in unregister_netdevice(), there is > a synchronize_rcu() before the netdev is actually unregistered and > freed. This ensures that any readers i.e. xmit() that got a non-NULL > peer will complete before the netdev is freed. > > Any readers after the RCU_INIT_POINTER() but before synchronize_rcu() > will dereference NULL, making it safe. > > The codepath to nsim_destroy() and nsim_create() takes both the newly > added nsim_dev_list_lock and rtnl_lock. This makes it safe with > concurrent calls to linking two netdevsims together. > > Signed-off-by: David Wei <dw@davidwei.uk> > --- > drivers/net/netdevsim/netdev.c | 27 ++++++++++++++++++++++----- > drivers/net/netdevsim/netdevsim.h | 1 + > 2 files changed, 23 insertions(+), 5 deletions(-) > > diff --git a/drivers/net/netdevsim/netdev.c b/drivers/net/netdevsim/netdev.c > index 9063f4f2971b..c3f3fda5fdc0 100644 > --- a/drivers/net/netdevsim/netdev.c > +++ b/drivers/net/netdevsim/netdev.c > @@ -29,18 +29,35 @@ > static netdev_tx_t nsim_start_xmit(struct sk_buff *skb, struct net_device *dev) > { > struct netdevsim *ns = netdev_priv(dev); > + unsigned int len = skb->len; > + struct netdevsim *peer_ns; > > + rcu_read_lock(); > if (!nsim_ipsec_tx(ns, skb)) > - goto out; > + goto out_drop_free; > > + peer_ns = rcu_dereference(ns->peer); > + if (!peer_ns) > + goto out_drop_free; > + > + skb_tx_timestamp(skb); > + if (unlikely(dev_forward_skb(peer_ns->netdev, skb) == NET_RX_DROP)) > + goto out_drop_cnt; > + > + rcu_read_unlock(); > u64_stats_update_begin(&ns->syncp); > ns->tx_packets++; > - ns->tx_bytes += skb->len; > + ns->tx_bytes += len; > u64_stats_update_end(&ns->syncp); > + return NETDEV_TX_OK; > > -out: > +out_drop_free: > dev_kfree_skb(skb); > - > +out_drop_cnt: > + rcu_read_unlock(); > + u64_stats_update_begin(&ns->syncp); > + ns->tx_dropped++; > + u64_stats_update_end(&ns->syncp); > return NETDEV_TX_OK; > } > > @@ -70,6 +87,7 @@ nsim_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats) > start = u64_stats_fetch_begin(&ns->syncp); > stats->tx_bytes = ns->tx_bytes; > stats->tx_packets = ns->tx_packets; > + stats->tx_dropped = ns->tx_dropped; > } while (u64_stats_fetch_retry(&ns->syncp, start)); > } > > @@ -302,7 +320,6 @@ static void nsim_setup(struct net_device *dev) > eth_hw_addr_random(dev); > > dev->tx_queue_len = 0; > - dev->flags |= IFF_NOARP; > dev->flags &= ~IFF_MULTICAST; > dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | > IFF_NO_QUEUE; > diff --git a/drivers/net/netdevsim/netdevsim.h b/drivers/net/netdevsim/netdevsim.h > index c8b45b0d955e..553c4b9b4f63 100644 > --- a/drivers/net/netdevsim/netdevsim.h > +++ b/drivers/net/netdevsim/netdevsim.h > @@ -98,6 +98,7 @@ struct netdevsim { > > u64 tx_packets; > u64 tx_bytes; > + u64 tx_dropped; > struct u64_stats_sync syncp; > > struct nsim_bus_dev *nsim_bus_dev; Reviewed-by: Maciek Machnikowski <maciek@machnikowski.net>
diff --git a/drivers/net/netdevsim/netdev.c b/drivers/net/netdevsim/netdev.c index 9063f4f2971b..c3f3fda5fdc0 100644 --- a/drivers/net/netdevsim/netdev.c +++ b/drivers/net/netdevsim/netdev.c @@ -29,18 +29,35 @@ static netdev_tx_t nsim_start_xmit(struct sk_buff *skb, struct net_device *dev) { struct netdevsim *ns = netdev_priv(dev); + unsigned int len = skb->len; + struct netdevsim *peer_ns; + rcu_read_lock(); if (!nsim_ipsec_tx(ns, skb)) - goto out; + goto out_drop_free; + peer_ns = rcu_dereference(ns->peer); + if (!peer_ns) + goto out_drop_free; + + skb_tx_timestamp(skb); + if (unlikely(dev_forward_skb(peer_ns->netdev, skb) == NET_RX_DROP)) + goto out_drop_cnt; + + rcu_read_unlock(); u64_stats_update_begin(&ns->syncp); ns->tx_packets++; - ns->tx_bytes += skb->len; + ns->tx_bytes += len; u64_stats_update_end(&ns->syncp); + return NETDEV_TX_OK; -out: +out_drop_free: dev_kfree_skb(skb); - +out_drop_cnt: + rcu_read_unlock(); + u64_stats_update_begin(&ns->syncp); + ns->tx_dropped++; + u64_stats_update_end(&ns->syncp); return NETDEV_TX_OK; } @@ -70,6 +87,7 @@ nsim_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats) start = u64_stats_fetch_begin(&ns->syncp); stats->tx_bytes = ns->tx_bytes; stats->tx_packets = ns->tx_packets; + stats->tx_dropped = ns->tx_dropped; } while (u64_stats_fetch_retry(&ns->syncp, start)); } @@ -302,7 +320,6 @@ static void nsim_setup(struct net_device *dev) eth_hw_addr_random(dev); dev->tx_queue_len = 0; - dev->flags |= IFF_NOARP; dev->flags &= ~IFF_MULTICAST; dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE; diff --git a/drivers/net/netdevsim/netdevsim.h b/drivers/net/netdevsim/netdevsim.h index c8b45b0d955e..553c4b9b4f63 100644 --- a/drivers/net/netdevsim/netdevsim.h +++ b/drivers/net/netdevsim/netdevsim.h @@ -98,6 +98,7 @@ struct netdevsim { u64 tx_packets; u64 tx_bytes; + u64 tx_dropped; struct u64_stats_sync syncp; struct nsim_bus_dev *nsim_bus_dev;
Forward skbs sent from one netdevsim port to its connected netdevsim port using dev_forward_skb, in a spirit similar to veth. Add a tx_dropped variable to struct netdevsim, tracking the number of skbs that could not be forwarded using dev_forward_skb(). The xmit() function accessing the peer ptr is protected by an RCU read critical section. The rcu_read_lock() is functionally redundant as since v5.0 all softirqs are implicitly RCU read critical sections; but it is useful for human readers. If another CPU is concurrently in nsim_destroy(), then it will first set the peer ptr to NULL. This does not affect any existing readers that dereferenced a non-NULL peer. Then, in unregister_netdevice(), there is a synchronize_rcu() before the netdev is actually unregistered and freed. This ensures that any readers i.e. xmit() that got a non-NULL peer will complete before the netdev is freed. Any readers after the RCU_INIT_POINTER() but before synchronize_rcu() will dereference NULL, making it safe. The codepath to nsim_destroy() and nsim_create() takes both the newly added nsim_dev_list_lock and rtnl_lock. This makes it safe with concurrent calls to linking two netdevsims together. Signed-off-by: David Wei <dw@davidwei.uk> --- drivers/net/netdevsim/netdev.c | 27 ++++++++++++++++++++++----- drivers/net/netdevsim/netdevsim.h | 1 + 2 files changed, 23 insertions(+), 5 deletions(-)